High-Fidelity 4x Neural Reconstruction of Real-time Path Traced Videos

Zhiqiang Lao, Zekai Chen, Jiali Cui, Marcos Conde, Radu Timofte, Heather Yu; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 5138-5146

The demand for high-fidelity, high-resolution video rendering in real-time applications presents significant challenges, particularly under low-sample path tracing where inputs are both noisy and low-resolution (LR). Conventional video super-resolution methods excel at spatial upsampling but assume clean inputs, while image-based denoising and super-resolution pipelines ignore temporal dependencies, often resulting in frame-level artifacts. We propose a unified deep learning framework for video reconstruction that integrates denoising and super-resolution into a temporally consistent pipeline. Specifically, our method first applies low-resolution denoising using paired noisy/clean data to ensure stable inputs, followed by a patch-level video upsampling stage for temporal super-resolution. To further enhance scene fidelity, we introduce a novel scene-level refinement stage that augments training with overlooked regions, enforcing consistency and coherence beyond patch-based sampling. This design effectively suppresses noise, recovers high-frequency details, and preserves temporal coherence. Experiments on challenging 4xupscaling tasks demonstrate that our method achieves high-quality video reconstructions with substantial fidelity improvements, while reducing rendering overhead in real-time.